---
license: cc-by-nc-4.0
language:
- en
pretty_name: RipBench
tags:
- rip currents
- rip_current
- beach safety
- computer vision
- classification
- object detection
- oriented object detection
- instance segmentation
- semantic segmentation
- panoptic segmentation
- video
size_categories:
- 10K<n<100K
---
# RipBench v1.2.2
This Readme describes the RipBench dataset, its contents, structure, splits, known limitations, how to use it and what to expect in future updates. For more details and other information, keep an eye on the [RipBench website](https://ripbench.ai) or write to andrei.dumitriu@uni-wuerzburg.de .

## Table of Contents
1. [Short Description](#short-description)
1. [Links of Interest](#links-of-interest)
1. [Structure](#structure)
1. [Tasks and Annotation Formats](#tasks-and-annotation-formats)
1. [Splits](#splits)
1. [Evaluation](#evaluation)
1. [Citations](#citations)
1. [Contributing](#contributing)
1. [Licensing](#licensing)
1. [Known Limitations](#known-limitations)
1. [Future Updates](#future-updates)
1. [Current Version](#current-version)

## Short description
RipBench was introduced with the paper **RipBench: A Unified Benchmark for Multi-Level Rip Current Detection, Classification and Segmentation**, accepted at [NeurIPS 2026](https://neurips.cc/). It is the result of a collaboration between the [Computer Vision Lab](https://www.informatik.uni-wuerzburg.de/computervision/) (CAIDAS & IFI) of the [University of Würzburg](https://www.uni-wuerzburg.de/en/) and the [University of Bucharest](https://unibuc.ro/).

RipBench asks the same beach footage six questions: is a rip current present (classification), where is it (axis-aligned detection, oriented detection), what is its exact extent (instance segmentation), and how does it sit in its surroundings (semantic and panoptic segmentation with the context classes water, land and sky). The dataset consists of 304 videos, evenly balanced between 152 videos with rip currents and 152 without, 303,491 frames in total, sampled in frame order at a minimum of 5 FPS into 51,011 annotated frames. Every annotated frame carries labels for all six tasks, produced by rip current experts and kept consistent across tasks. The train split additionally offers, as an opt-in scenario, 3,773 standalone stills annotated for boxes, oriented boxes and instance masks.

RipBench builds on the video pool of [RipVIS](https://huggingface.co/datasets/Irikos/RipVIS) (all 150 rip videos and 33 of its 34 non-rip videos are inherited and were re-annotated) and adds newly collected footage. It is authored by:
- Andrei Dumitriu (andrei.dumitriu@uni-wuerzburg.de)
- Aakash Ralhan
- Conf. Dr. Florin Tatui
- Florin Miron
- Prof. Dr. Radu Tudor Ionescu
- Prof. Dr. Radu Timofte

**[⬆ back to top](#table-of-contents)**

## Links of interest
- Main contact: andrei.dumitriu@uni-wuerzburg.de
- Website: https://ripbench.ai
- HuggingFace link: https://huggingface.co/datasets/Irikos/RipBench
- Previous dataset: https://huggingface.co/datasets/Irikos/RipVIS
- Evaluation on the test split: see [Evaluation](#evaluation)

**[⬆ back to top](#table-of-contents)**

## Structure
This upload is the minimal, self-contained version of RipBench v1.2.2: every video, every sampled frame, the annotations of the train and validation splits for all six tasks, the split lists, the evaluators and the per-video metadata. Larger optional parts (see [Future Updates](#future-updates)) follow in later uploads.

```
RipBench_v1.2.2
    - videos/
        -- rips/                       the 152 rip videos, RipBench-<video>.mp4
        -- no-rips/                    the 152 non-rip videos, RipBench-NR-<video>.mp4
    - images/
        -- sampled_frames/rips/        frames of the 152 rip videos, RipBench-<video>_<frame>.jpg
        -- sampled_frames/no-rips/     frames of the 152 non-rip videos, RipBench-NR-<video>_<frame>.jpg
        -- additional_data/rips/       2,466 standalone stills with rip currents (opt-in train scenario)
        -- additional_data/no-rips/    1,307 standalone stills without rip currents (opt-in train scenario)
    - splits/s1.3/
        -- train.txt, val.txt, test.txt   one image path per line; train.txt includes the additional stills
    - labels/                           training-ready views, one folder per task (see labels/README.md)
        -- classification/              train / train_with_additional_data / val manifests: `image_path class_id`
        -- bbox/{yolo26,coco}/           axis-aligned boxes (manual tight-box pass)
        -- bbox_from_instance/{yolo26,coco}/   boxes derived from the instance masks (box-protocol ablation)
        -- obb/{yolo26,dota}/            oriented boxes
        -- instance_segmentation/{yolo26,coco}/   instance masks
        -- semantic_segmentation/       masks/ (grayscale PNG, 1 rip_current, 2 water, 3 land, 4 sky) + split lists
        -- panoptic_segmentation/coco/  COCO panoptic jsons + RGB segment-id PNGs
        -- video_instance_segmentation/ytvis_sparse/   YouTube-VIS json with reviewed track ids (train, val)
    - tooling/eval/                     the released evaluators (classification, detection, OBB, semantic) and helpers
    - video_metadata.csv                one row per video: class, split, origin, source tag, licence, resolution, fps, counts
    - dataset_summary_by_video.csv      per-video frame counts
    - metadata/item_metadata.csv        per-item sub-licences and attribution, where they differ from the base licence
    - provenance/                       how the attribution was derived; people audit; annotation-revision notes
    - RipBench_version_changelog.md     every dataset version and what changed
    - LICENSE, checksums.sha256 (extracted files), videos.sha256, ARCHIVES.sha256, README.md (this file)
```

### Archives on Hugging Face
The large folders are uploaded as archives created from the dataset root, so extracting all of them in place recreates exactly the tree above (`tar -xf <archive>` in the folder that holds `README.md`); `checksums.sha256` lists every extracted file and `ARCHIVES.sha256` the archives themselves.

| archive | unpacks to | content |
|---|---|---|
| `images/sampled_frames_rips.tar` | `images/sampled_frames/rips/` | frames of the 152 rip videos |
| `images/sampled_frames_no-rips.tar` | `images/sampled_frames/no-rips/` | frames of the 152 non-rip videos |
| `images/additional_data.tar` | `images/additional_data/` | the 3,773 standalone stills |
| `labels/classification.tar.gz`, `labels/bbox.tar.gz`, `labels/bbox_from_instance.tar.gz`, `labels/obb.tar.gz`, `labels/instance_segmentation.tar.gz` | `labels/<task>/` | text-format annotations of that task |
| `labels/semantic_segmentation.tar`, `labels/panoptic_segmentation.tar` | `labels/<task>/` | PNG masks and jsons of that task |
| `labels/video_instance_segmentation.tar.gz` | `labels/video_instance_segmentation/` | the sparse YouTube-VIS view |

Everything else (`videos/`, `splits/`, `tooling/`, `provenance/`, `metadata/`, the csv files, `LICENSE`, the changelog and this readme) is uploaded as plain files; the videos are plain MP4 files, one per video, listed with their checksums in `videos.sha256`.

Frames of a video are named `RipBench-<video>_<frame>.jpg` (rip videos) or `RipBench-NR-<video>_<frame>.jpg` (NR = no rips); the video number is zero-padded to 3 digits and the frame number to 5 digits (the 28th frame of the 4th video is `RipBench-004_00028.jpg`). Results are reported per frame.

**[⬆ back to top](#table-of-contents)**

## Tasks and Annotation Formats
| Task | Where | Format |
|---|---|---|
| Classification (rip present / absent) | `labels/classification/` | manifests with one `image_path class_id` line per frame; the label is per frame (rip = 1 iff the frame has a box) |
| Axis-aligned detection | `labels/bbox/` | YOLO txt (normalized xywh) and COCO json; a separate manual annotation pass |
| Boxes derived from masks | `labels/bbox_from_instance/` | same layouts; bounding rectangles of the instance polygons |
| Oriented detection | `labels/obb/` | YOLO-OBB txt (8 normalized corner values) and DOTA txt (pixel coordinates) |
| Instance segmentation | `labels/instance_segmentation/` | YOLO polygon txt and COCO json |
| Semantic segmentation | `labels/semantic_segmentation/` | grayscale PNG masks, pixel values 1 rip_current / 2 water / 3 land / 4 sky |
| Panoptic segmentation | `labels/panoptic_segmentation/` | COCO panoptic json + RGB-encoded segment-id PNGs, rip_current the only thing class |
| Video instance segmentation | `labels/video_instance_segmentation/ytvis_sparse/` | YouTube-VIS-2019 json, annotated frames as low-fps sequences with reviewed track identities |

The single foreground class is `rip_current` (id 0 in YOLO formats, id 1 in COCO jsons). Semantic and panoptic labels exist for the sampled video frames; the standalone stills are annotated for boxes, oriented boxes and instance masks only. All paths inside `labels/` are relative, so copying `images/`, `labels/` and `splits/` to another machine needs no rewriting.

**[⬆ back to top](#table-of-contents)**

## Splits
The split (`s1.3`) is video-level and was assigned manually by the rip current experts of the team, balancing video durations, annotated-frame counts, viewpoints and rip current types, so that no frames of one video appear in two splits. It is the split every number in the paper is computed on. Use it as is: automatic splits of videos, or worse of frames, leak near-identical frames between splits and overstate results.

| split | rip videos | non-rip videos | annotated frames |
|---|---|---|---|
| train | 91 | 74 | 31,083 (34,856 with the additional stills) |
| val | 30 | 38 | 10,369 |
| test | 31 | 40 | 9,559 |

### Train
RipBench-
```['003', '004', '005', '006', '010', '011', '016', '017', '018', '020', '021', '022', '028', '029', '030', '031', '032', '033', '034', '035', '036', '037', '040', '041', '042', '045', '049', '050', '051', '052', '054', '056', '057', '058', '060', '061', '062', '063', '064', '065', '068', '069', '070', '071', '075', '076', '077', '080', '082', '083', '084', '085', '086', '087', '088', '089', '091', '092', '093', '094', '095', '096', '097', '098', '100', '101', '103', '104', '105', '106', '107', '110', '111', '112', '115', '116', '117', '118', '122', '123', '124', '126', '127', '135', '136', '140', '141', '148', '149', '150', '152']```

RipBench-NR-
```['002', '003', '006', '007', '008', '010', '011', '012', '015', '016', '017', '018', '025', '027', '028', '029', '030', '031', '032', '033', '034', '035', '039', '040', '041', '042', '044', '048', '049', '050', '057', '058', '061', '064', '066', '068', '069', '070', '071', '075', '076', '078', '084', '085', '091', '097', '098', '099', '100', '103', '105', '108', '109', '110', '111', '112', '115', '117', '122', '123', '124', '125', '130', '131', '133', '135', '136', '138', '139', '140', '143', '146', '147', '149']```

### Validation
RipBench-
```['001', '007', '012', '014', '015', '019', '023', '024', '026', '039', '044', '046', '047', '053', '066', '067', '081', '090', '102', '108', '109', '121', '128', '129', '133', '134', '137', '143', '144', '146']```

RipBench-NR-
```['004', '009', '019', '020', '022', '024', '026', '037', '043', '045', '046', '051', '053', '054', '056', '059', '062', '063', '067', '072', '080', '081', '082', '086', '087', '089', '090', '096', '101', '102', '107', '114', '118', '126', '132', '144', '148', '150']```

### Test
RipBench-
```['002', '008', '009', '013', '025', '027', '038', '043', '048', '055', '059', '072', '073', '074', '078', '079', '099', '113', '114', '119', '120', '125', '130', '131', '132', '138', '139', '142', '145', '147', '151']```

RipBench-NR-
```['001', '005', '013', '014', '021', '023', '036', '038', '047', '052', '055', '060', '065', '073', '074', '077', '079', '083', '088', '092', '093', '094', '095', '104', '106', '113', '116', '119', '120', '121', '127', '128', '129', '134', '137', '141', '142', '145', '151', '152']```

**[⬆ back to top](#table-of-contents)**

## Evaluation
The test split is held out: its frames and its list (`splits/s1.3/test.txt`) are included, its annotations are not, for any task. Train and validate on the released labels, select every operating point (confidence threshold, checkpoint) on the validation split, and submit test-split predictions through the benchmark's evaluation procedure announced on https://ripbench.ai. The released evaluators in `tooling/eval/` implement the paper's protocol (per-frame scoring, operating thresholds frozen on validation, $F_2$ as the headline metric because a missed rip current costs more than a false alarm) and can be run locally on the validation split.

**[⬆ back to top](#table-of-contents)**

## Citations
If you use RipBench, please cite:

1. **RipBench: A Unified Benchmark for Multi-Level Rip Current Detection, Classification and Segmentation**
<pre>
@inproceedings{dumitriu2026ripbench,
    author    = {Dumitriu, Andrei and Ralhan, Aakash and Tatui, Florin and Miron, Florin and Ionescu, Radu Tudor and Timofte, Radu},
    title     = {RipBench: A Unified Benchmark for Multi-Level Rip Current Detection, Classification and Segmentation},
    booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
    year      = {2026}
}
</pre>

The video pool and the standalone stills come from our earlier work, which the dataset also builds on:

2. **RipVIS: Rip Currents Video Instance Segmentation Benchmark for Beach Monitoring and Safety**
<pre>
@inproceedings{dumitriu2025ripvis,
    author    = {Dumitriu, Andrei and Tatui, Florin and Miron, Florin and Ralhan, Aakash and Ionescu, Radu Tudor and Timofte, Radu},
    title     = {RipVIS: Rip Currents Video Instance Segmentation Benchmark for Beach Monitoring and Safety},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    year      = {2025},
    pages     = {3427--3437}
}
</pre>

3. **Rip Current Segmentation: A Novel Benchmark and YOLOv8 Baseline Results**
<pre>
@inproceedings{dumitriu2023rip,
    title     = {Rip Current Segmentation: A Novel Benchmark and YOLOv8 Baseline Results},
    author    = {Dumitriu, Andrei and Tatui, Florin and Miron, Florin and Ionescu, Radu Tudor and Timofte, Radu},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
    pages     = {1261--1271},
    year      = {2023}
}
</pre>

**[⬆ back to top](#table-of-contents)**

## Contributing
We welcome contributions; see https://ripbench.ai or write to the main contact, including suggestions for this readme.

### Contributing footage
RipBench keeps growing. If you have a video with rip currents (or beach footage without them, which is just as useful), send it to us: we annotate it for all six tasks and include it in a future version, under a licence decided by you and with the source credited to you.

**[⬆ back to top](#table-of-contents)**

## Licensing
RipBench is released under the Creative Commons Attribution-NonCommercial 4.0 International License ([CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/)): videos, stills and all annotations. Where an item carries a different sub-licence or a required attribution, it is listed in `metadata/item_metadata.csv`, and the origin and source tag of every video is in `video_metadata.csv`.

In summary:
1. Non-commercial use only
2. Attribution required

By downloading or using RipBench, you agree to these terms. For inquiries regarding commercial use, please contact the main author (andrei.dumitriu@uni-wuerzburg.de).

**[⬆ back to top](#table-of-contents)**

## Known Limitations
1. The frame-level class of a test frame can be read off the folder it sits in (`rips/` vs `no-rips/`) and the video id (`RipBench-NR-`). The withheld test annotations are the spatial ones (boxes, oriented boxes, masks, panoptic segments) and the per-frame presence labels; do not use the folder or the id as a prediction.
1. 344 frames at the tail of RipBench-044 contain no rip current (the drone moves away mid-video) and are labelled accordingly in every task, although the video sits in the `rips/` folder; presence labels are per frame, not per video.
1. RipBench-152 is the only flash-rip video; it is in the train split, so the benchmark does not evaluate flash-rip generalization.
1. The numbers in the paper were computed on dataset version 1.2.0. Version 1.2.1 refreshed the water/land/sky classes of 9,348 semantic and panoptic masks (rip pixels and segment ids unchanged); the rip-class columns of the paper are therefore identical on this release, the background and overall panoptic/semantic columns are not reproducible from it. Details in `provenance/background_refresh_v1.2.1.md`.
1. The standalone stills (`additional_data/`) are annotated for boxes, oriented boxes and instance masks only; they have no semantic or panoptic labels and appear only in the train split.
1. People appear in the footage (public beaches) and were not blurred; see `provenance/people_audit_videos.tsv` and `people_audit_frames.tsv`.

**[⬆ back to top](#table-of-contents)**

## Future Updates
1. The source annotations (`labels_original/`) and the label builders, so that every view can be regenerated.
1. The derived video views (densified and curated YouTube-VIS views, task-track views) and the video-model baselines.
1. The paper's result files, the baseline checkpoints and a results website with per-model provenance.
1. A Croissant metadata record for this release.
1. The evaluation procedure for the test split.

**[⬆ back to top](#table-of-contents)**

## Current Version
Current version of RipBench is 1.2.2 (public; internal versions carry odd patch numbers). See `RipBench_version_changelog.md` for every version.

Last DATASET update: 06.11.2026
Last README update: 06.11.2026

**[⬆ back to top](#table-of-contents)**